Bartel enrichMiR analysis 2

suppressPackageStartupMessages({
  library(SummarizedExperiment)
  library(SEtools)
  library(ggplot2)
  library(plgINS)
  library(edgeR)
  library(DT)
  library(viridis)
  library(RColorBrewer)
  library(data.table)
  library(dplyr)
  library(tibble)
  library(cowplot)
  library(matrixStats)
  library(RUVSeq)
})

## Warning: replacing previous import 'GenomicRanges::shift' by 'data.table::shift'
## when loading 'plgINS'

Allocation

# input file: Bartel DEA SE
input <- "data/bartel.hela.DEA.SE.rds"
#input <- "data/bartel.hek.DEA.SE.rds"

# output file: enrichMiR results
output.e <- "data/bartel.hela.e.TSperm.rds"
#output.e <- "data/bartel.hek.e.TSperm.rds"

# number of cores for high runtime functions
cores <- 5

# set number of replicates per permutation proportion
nrep <- 3

Loading

# import
se <- readRDS(input)

Permutations

# define variables
#treatment <- unique(as.character(se$miRNA))
dea.df.names <- colnames(rowData(se)[,grepl("DEA",colnames(rowData(se)))][-2])
dea.names <- gsub("DEA.","", dea.df.names)

# save Bartel DEAs into list
dea.list <- lapply(rowData(se)[,dea.df.names], function(x) x)

enrichMiR

# load enrichMiR package
devtools::load_all("/mnt/schratt/tgermade_test/master_19_20/enrichMiR/enrichMiR/")

#' getDEA
#'
#' @param dea.df dataframe of DEA results (an SE rowData column if generated via DEA() function)
#'
#' @return FDR-filtered dataframe containing c("symbol","logFC","PValue","FDR") columns
#'
getDEA <- function(dea.df){
  if(!any(colnames(dea.df) %in% "symbol")) dea.df$symbol <- rownames(dea.df)
  dea <- as.data.frame(dea.df[,c("symbol","logFC", "logCPM","PValue","FDR")])
  dea <- aggregate(dea[,-1], by=list(symbol=dea$symbol),FUN=mean)
  rownames(dea) <- dea$symbol
  dea <- dea[,-1]
  
  lapply(dea, FUN = function(x){
    if(any(is.infinite(x))){
      w <- which(is.infinite(x))
      x[w] <- max(abs(x[setdiff(which(dea$FDR<0.5),w)]))*sign(x[w])
    }
  })
  return(dea)
}

# generate dea object (high runtime!)
#dea.list <- lapply(dea.df.names, FUN = function(x) getDEA(rowData(se.hela)[[x]]))

dea.list <- bplapply(dea.list, getDEA, 
                     BPPARAM = MulticoreParam(cores, progressbar = FALSE) )
names(dea.list) <- dea.names

# we don't use miRNA expression values for the benchmark. This will make it more difficult # for the enrichMiR tests
mirexpr <- NULL

#' getTS
#'
#' @param species character object. Can be "human", "mouse" or "rat"
#'
#' @return TargetScan miRNA target dataframe with family information in metadata()
#'
getTS <- function(species=c("human","mouse","rat")){
  library(S4Vectors)
  species <- match.arg(species)
  
  # assign species ID
  spec <- switch( species,
                  human = 9606,
                  mouse = 10090,
                  rat = 10116 )
  
  
  # download TargetScan miRNA targeting dataset
  tmp <- tempfile()
  download.file(
    "http://www.targetscan.org/vert_72/vert_72_data_download/Summary_Counts.all_predictions.txt.zip", 
    tmp)
  unzip(tmp)
  full <- fread("Summary_Counts.all_predictions.txt") #, sep = "\t", header = TRUE)
  
  # limit to selected species
  sub <- full[full$'Species ID' == spec,]
  sub$score <- as.numeric(as.character(sub$'Cumulative weighted context++ score'))
    
  # generate TargetScan dataframe
  ts <- DataFrame(family = sub$'miRNA family',
                   rep.miRNA = sub$'Representative miRNA',
                   feature = sub$'Gene Symbol',
                   sites = sub$'Total num conserved sites',
                   score = as.numeric(as.character(sub$'Cumulative weighted context++ score'))
                   )
  ts <- DataFrame(
    aggregate(sub[,c("sites","score")], by=ts[,c("family","feature")], FUN=mean)
    )
  

  
  # download TargetScan miRNA families dataset
  tmp <- tempfile()
  download.file(
    "http://www.targetscan.org/vert_72/vert_72_data_download/miR_Family_Info.txt.zip", 
    tmp)
  unzip(tmp)
  full <- fread("miR_Family_Info.txt") #, sep = "\t", header = TRUE)
  
  # limit to selected species
  sub <- full[full$'Species ID' == spec,]
  
  fam <- sub$`Seed+m8`
  names(fam) <- sub$`MiRBase ID`
  
  # add family info to ts dataframe as attribute
  metadata(ts)$families <- fam
  
  # enrichMiR cant handle 0 values for sites feature
  return(ts[ts$sites!=0,])
}

# get TargetScan miRNA targets
TS <- getTS(species = "human")

# load TS object
TS <- readRDS( "data/TargetScan_all.rds")

# load detailed Bartel treatment info containing exact sequences
pert <- read.csv("data/bartel_treatments.txt", sep = "\t")

# we extract HeLa & HEK data (we disregard passenger strands as long as we're working with TargetScan)
colnames(pert) <- c("treatment","seq")
pert <- list(pert[1:17,], pert[37:48,])
names(pert) <- c("hela","hek")

if(grepl("hela", input)){
  pert <- pert$hela
} else if(grepl("hek", input)){
  pert <- pert$hek
}

# find out where the seed sequence begins
#unlist(gregexpr(pattern ="GAGGUAG",pert$seq))
#unlist(gregexpr(pattern ="GGAAUGU",pert$seq))

# extract seed sequences
pert$seed <- sapply(pert$seq, function(x) substring(x ,2, 8))

# get seed family for each treatment miRNA (true positive)
TP <- sapply(dea.names, FUN=function(x) unique(as.character(
  pert$seed[grepl(paste0(x,"\\("), pert$treatment) | grepl(paste0(x,"$"), pert$treatment)] 
  )) )

#' TSperm
#'
#' @param TS TargetScan DataFrame of miRNA tx targets
#' @param genes string vector of gene names
#' @param props vector of proportions, e.g. c(.2,.3,.4,.5)
#'
#' @return TS DataFrame with permuted miRNA targets
#' 
TSperm <- function(TS, TP, genes, props){
  # get TargetScan data for each treatment miRNA family
  TS.part <- TS[TS$family==TP,]
  lapply(props, FUN=function(p){
    sn <- floor(p*nrow(TS.part))
    genes <- genes[!genes %in% TS.part$feature]
    i <- sample(seq_len(nrow(TS.part)), sn)
    j <- sample(seq_len(length(genes)), sn)
    TS.part$feature[i] <- as.character(genes[j])
    TS <- rbind(TS.part, TS[TS$family!=TP,])
    return(TS)
  })
}

# the permuatation proportions that should be computed for each DEA
names(i) <- i <- 1:nrep
props <- unlist(lapply( c("20"=0.2, "30"=0.3, "40"=0.4, "50"=0.5), 
                        FUN=function(x) lapply(i, FUN=function(y) x)))
# do TS permutations
set.seed(1234)
TS.list <- bplapply(dea.names, FUN=function(i){
  TSperm(TS, TP[i], rownames(se), props)
}, BPPARAM = MulticoreParam(cores, progressbar = FALSE) )

# naming
names(TS.list) <- dea.names
for(i in dea.names){
  names(TS.list[[i]]) <- names(props)
}
# add original TS to permutated ones
for(i in dea.names){
  TS.list[[i]][["original"]] <- TS
}
# place original at first place
for(i in dea.names){
  TS.list[[i]] <- TS.list[[i]][c("original", names(props))]
}

props.all <- c(original="original", props)

tests <- c("areamir","overlap","michael","KS","KS2","MW")
# all tests except WO:
tests <- c("overlap","michael","mw","ks","ks2","areamir","areamir2")
# all tests
tests <- c("overlap","michael","wo","mw","ks","ks2","gsea","gseamod","modscore","modsites","areamir","areamir2")
tests <- NULL
cores <- 8

# run enrichMiR on all objects of dea.list (high runtime!)

## foreach: doesnt work
library(foreach)
library(doParallel)
# cl <- makeCluster(cores)
# registerDoParallel(cl, cores=cores)
# chunk.size <- length(dea.names)/cores
# test <- foreach(c=1:cores, .combine="rbind") %dopar% {
#   for(i in dea.names[((c-1)*chunk.size+1):(c*chunk.size)]){
#     lapply(names(props.all), FUN=function(j){
#       enrichMiR(DEA=as.data.frame(dea.list[[i]]), TS=TS.list[[i]][[j]], miRNA.expression=mirexpr, 
#                 cleanNames=TRUE, tests=tests)
#       })
#     }
# }
#stopImplicitCluster()
# stopCluster(cl)

## foreach: Pierre's code (works, but 1 core)
res <- foreach( dea=dea.list, 
                TS=unlist(TS.list,recursive=FALSE) ) %dopar% {
                  enrichMiR( as.data.frame(dea), TS=TS, miRNA.expression=mirexpr, 
                             cleanNames=TRUE, tests=tests )
                }

## mclapply (works, but 1 core; all cores when reduced tests var)
## looks like GSEA, WO, mods are problematic
library(parallel)
# multicores on Linux
e.list <- mclapply(dea.names, FUN=function(i){
  lapply(names(props.all), FUN=function(j){
      enrichMiR(DEA=as.data.frame(dea.list[[i]]), TS=TS.list[[i]][[j]], miRNA.expression=mirexpr, 
                cleanNames=TRUE, tests=tests)
      })
  }, mc.cores = cores )
## mcmapply: inspired by Pierre's code
res <- mcmapply(dea=dea.list, 
                TS=unlist(TS.list,recursive=FALSE), 
                FUN=function(x){
                  enrichMiR( as.data.frame(dea), TS=TS, miRNA.expression=mirexpr, 
                             cleanNames=TRUE, tests=tests )
                  }, mc.cores = cores )


## bplapply: doesnt work at all
e.list <- bplapply(dea.names[1:4], FUN=function(i){
  lapply(names(props.all), FUN=function(j){
      enrichMiR(DEA=as.data.frame(dea.list[[i]]), TS=TS.list[[i]][[j]], miRNA.expression=mirexpr,
                cleanNames=TRUE, tests=tests)
      })
  }, BPPARAM = MulticoreParam(cores, progressbar = TRUE) )


## bpmapply: Pierre's code
res <- bpmapply(dea=dea.list, 
                TS=unlist(TS.list,recursive=FALSE), 
                BPPARAM=MulticoreParam(4, threshold="TRACE"), FUN=function(x){
                  enrichMiR( as.data.frame(dea), TS=TS, miRNA.expression=mirexpr, cleanNames=TRUE, tests=tests )
                  } )


## serial run
e.list <- lapply(dea.names, FUN=function(i){
  lapply(names(props.all), FUN=function(j){
      enrichMiR(DEA=as.data.frame(dea.list[[i]]), TS=TS.list[[i]][[j]], miRNA.expression=mirexpr, 
                cleanNames=TRUE, tests=tests)
      })
  })


# naming
names(e.list) <- dea.names
for(i in dea.names){
  names(e.list[[i]]) <- names(props.all)
}

# use this between parallel runs
gc(verbose = T)

# save e object
saveRDS(e.list, output.e)

# load TS object
e.list <- readRDS(output.e)

Benchmarking

dat <- function(e, thresholds, TP){
  dplyr::bind_rows(
    lapply(e@res, FUN=function(x){
      if("family" %in% colnames(x)) row.names(x) <- x$family
      x$truth <- row.names(x) %in% TP
      x$FDR[is.na(x$FDR)] <- 1
      as.data.frame(
        t(sapply(thresholds, FUN=function(i){
          P <- sum(x$FDR<=i)
          c( threshold=i,
             P=P,
             TP=sum(x$FDR<=i & x$truth),
             FP=sum(x$FDR<=i & !x$truth),
             TPR=sum(x$FDR<=i & x$truth)/sum(x$truth),
             FPR=sum(x$FDR<=i & !x$truth)/sum(!x$truth),
             FDR=ifelse(P>0,sum(x$FDR<=i & !x$truth)/P,0) )
          }))
        )
      }), .id="method")
}

# Benchmarking

devtools::load_all("../enrichMiR/enrichMiR/")
fams <- metadata(TS)$families

#dea <- deas[[1]]
#dea <- rowData(se.hela)$`DEA.let-7a`
#e <- enrichMiR(dea, TS)
#e <- e.list$`miR-122`

#TP <- unique(as.character(fams[grep("miR-144$|miR-144-",names(fams))]))
thresholds <- c(0,10^(-10:-3),0.005,0.01,0.025,0.05,0.075,0.1,0.15,0.2,0.25,0.3,0.5)

# get benchmark results
dat.list <- lapply(dea.names[1], FUN=function(x) lapply(names(e.list[[x]]), FUN=function(y) dat(e.list[[x]][[y]], thresholds, TP.list[[x]]) ))

ggplot(dat.list[[1]][[2]], aes(FPR, TPR, colour=method)) + geom_line() + geom_point(size=3) + 
  scale_x_log10()  
#+ scale_colour_manual(values=cols)

#' doBenchmark
#'
#' @param res enrichMiR test results (e object)
#' @param TP character vector: contains the seeds (families) for a miRNA treatment
#'
#' @return a dataframe containing scores for each enrichMiR test
#'
doBenchmark <- function(res, TP){
  res <- lapply(res, FUN=function(x){
    x <- x[order(x$FDR),]
    if("family" %in% colnames(x)) row.names(x) <- x$family
    x$truth <- row.names(x) %in% TP
    x$FDR[is.na(x$FDR)] <- 1
    x
  })
  data.frame( method=names(res),
              detPPV = sapply(res, FUN=function(x) 1/which(x$truth)[1] ),
              FP.atFDR05 = sapply(res, FUN=function(x) sum(!x$truth & x$FDR<0.05)),
              log10QDiff = sapply(res, FUN=function(x){
                tp1 <- -log10(x$FDR[which(x$truth)[1]])
                fp1 <- -log10(x$FDR[which(!x$truth)[1]])
                tp1-fp1
              }),
              log10QrelIncrease = sapply(res, FUN=function(x){
                tp1 <- -log10(x$FDR[which(x$truth)[1]])
                fp1 <- -log10(x$FDR[which(!x$truth)[1]])
                (tp1-fp1)/(min(tp1,fp1))
              }),
              TP.atFDR05 = sapply(res, FUN=function(x) sum(x$truth[x$FDR<0.05]))
              )
  
}

# generate the benchmarking scores
BM.list <- lapply(dea.names, FUN=function(x){
  lapply(names(e.list[[x]]), FUN=function(y){
    doBenchmark(e.list[[x]][[y]]@res, TP[x]) 
    })
})
# naming
names(BM.list) <- dea.names
for(x in dea.names){
  names(BM.list[[x]]) <- names(e.list[[x]])
}

# generate a results df for plotting
BM.list2 <- lapply(BM.list, FUN=function(x) dplyr::bind_rows(x, .id = "prop.rep"))
BM.df <- dplyr::bind_rows(BM.list2, .id="treatment")
BM.df$prop <- unlist(lapply(strsplit(BM.df$prop.rep, "[.]"), FUN=function(x) x[1]))
BM.df$prop <- factor(BM.df$prop, levels=unique(BM.df$prop))

df <- BM.df

Plots

# plots
for(i in c("detPPV","FP.atFDR05","log10QDiff","TP.atFDR05")){
  cat("Score analysis: ", i,"\n\n")
  ## boxplot
  print( 
    ggplot(df, aes(x=prop, y=df[[i]], fill=method)) + geom_boxplot() + ylab(i) 
    )
  if(i!="TP.atFDR05"){
    ## density plot: score distribution per permutation
    print( 
      ggplot(df, aes(log10(df[[i]]))) + geom_density(aes(fill=method), alpha=0.9)  + facet_wrap(~prop) + xlab(i)
      )
    ## violin plot
    print( 
      ggplot(df, aes(x=prop, y=df[[i]], fill=method))  + geom_violin() + facet_wrap(~method) +
        guides(fill=FALSE) + ylab(i) + geom_jitter(shape=16, position=position_jitter(0.2))
      )
  }
  ## violin plot median
  print(
    ggplot(df, aes(x=prop, y=df[[i]], fill=method))  + geom_violin(draw_quantiles = 0.5) + facet_wrap(~method) +
      guides(fill=FALSE) + ylab(i)
    )
}

## Score analysis:  detPPV

## Score analysis:  FP.atFDR05

## Score analysis:  log10QDiff

## Score analysis:  TP.atFDR05

scores over the datasets

# mean scores over all datasets; this shows which datasets were difficult to handle for the tests
ggplot(df, aes(x=prop, y=detPPV, fill=treatment))  + geom_violin(draw_quantiles = 0.5) + facet_wrap(~treatment) +
      guides(fill=FALSE)

mean.datasets <- lapply( unique(df$treatment), FUN=function(x) 
  sapply( unique(df$prop), function(y) 
    round( mean(df$detPPV[df$treatment==x & df$prop==y], na.rm=TRUE), 3) 
    )
  )
names(mean.datasets) <- dea.names
for(i in dea.names){
  names(mean.datasets[[i]]) <- unique(df$prop)
}
mean.datasets

## $`let-7a`
## original       20       30       40       50 
##    0.768    0.768    0.768    0.768    0.645 
## 
## $`miR-1`
## original       20       30       40       50 
##    0.801    0.801    0.801    0.801    0.801 
## 
## $`miR-124`
## original       20       30       40       50 
##    0.801    0.801    0.779    0.539    0.385 
## 
## $`miR-137`
## original       20       30       40       50 
##    0.768    0.768    0.768    0.768    0.601 
## 
## $`miR-139`
## original       20       30       40       50 
##    0.411    0.356    0.229    0.260    0.094 
## 
## $`miR-143`
## original       20       30       40       50 
##    0.824    0.824    0.825    0.824    0.824 
## 
## $`miR-144`
## original       20       30       40       50 
##    0.825    0.641    0.426    0.296    0.211 
## 
## $`miR-153`
## original       20       30       40       50 
##    0.768    0.745    0.568    0.434    0.393 
## 
## $`miR-155`
## original       20       30       40       50 
##    0.795    0.782    0.778    0.781    0.777 
## 
## $`miR-182`
## original       20       30       40       50 
##    0.768    0.439    0.403    0.380    0.342 
## 
## $`miR-199a`
## original       20       30       40       50 
##    0.824    0.783    0.781    0.731    0.602 
## 
## $`miR-204`
## original       20       30       40       50 
##    0.795    0.779    0.779    0.745    0.705 
## 
## $`miR-205`
## original       20       30       40       50 
##    0.824    0.773    0.769    0.690    0.604 
## 
## $`miR-216b`
## original       20       30       40       50 
##    0.712    0.475    0.389    0.174    0.089 
## 
## $`miR-223`
## original       20       30       40       50 
##    0.824    0.824    0.824    0.815    0.759 
## 
## $`miR-7`
## original       20       30       40       50 
##    0.824    0.811    0.805    0.798    0.809

scores over methods

# mean scores over all methods
ggplot(df, aes(x=prop, y=detPPV, fill=method))  + geom_violin(draw_quantiles = 0.5) + facet_wrap(~method) +
      guides(fill=FALSE)

mean.methods <- lapply( unique(df$method), FUN=function(x) 
  sapply( unique(df$prop), function(y) 
    round( mean(df$detPPV[df$method==x & df$prop==y], na.rm=TRUE), 3)
    )
  )
names(mean.methods) <- unique(df$method)
for(i in unique(df$method)){
  names(mean.methods[[i]]) <- unique(df$prop)
}
mean.methods

## $aREAmir
## original       20       30       40       50 
##    0.969    0.849    0.768    0.788    0.571 
## 
## $EN.up
## original       20       30       40       50 
##    0.005    0.005    0.005    0.005    0.005 
## 
## $EN.down
## original       20       30       40       50 
##    1.000    0.919    0.828    0.773    0.663 
## 
## $EN.combined
## original       20       30       40       50 
##    0.950    0.863    0.803    0.730    0.590 
## 
## $wEN.up
## original       20       30       40       50 
##    0.005    0.005    0.005    0.005    0.005 
## 
## $wEN.down
## original       20       30       40       50 
##    1.000    0.938    0.906    0.820    0.768 
## 
## $wEN.combined
## original       20       30       40       50 
##    1.000    0.948    0.887    0.789    0.757 
## 
## $michael.up
## original       20       30       40       50 
##    0.005    0.005    0.005    0.005    0.005 
## 
## $michael.down
## original       20       30       40       50 
##    1.000    0.917    0.840    0.770    0.669 
## 
## $michael.combined
## original       20       30       40       50 
##    1.000    0.917    0.865    0.835    0.752 
## 
## $MW
## original       20       30       40       50 
##    0.939    0.841    0.779    0.677    0.558 
## 
## $KS
## original       20       30       40       50 
##    0.847    0.738    0.708    0.632    0.534 
## 
## $KS2
## original       20       30       40       50 
##    0.958    0.919    0.926    0.781    0.715 
## 
## $GSEA
## original       20       30       40       50 
##    0.832    0.779    0.732    0.668    0.701 
## 
## $GSEAmodified
## original       20       30       40       50 
##    0.704    0.603    0.541    0.511    0.474 
## 
## $modSites
## original       20       30       40       50 
##    0.938    0.896    0.830    0.772    0.699 
## 
## $modScore
## original       20       30       40       50 
##    0.943    0.919    0.897    0.837    0.737

TP ratio at FDR 0.05 over methods

# sensitivity score (ratio of TP at FDR 0.05)
sens.methods <- lapply( unique(df$method), FUN=function(x) 
  sapply( unique(df$prop), function(y)
    round( sum(df$TP.atFDR05==1 & df$method==x & df$prop==y) / sum(df$method==x & df$prop==y), 3)
    )
  )
names(sens.methods) <- unique(df$method)
for(i in unique(df$method)){
  names(sens.methods[[i]]) <- unique(df$prop)
}
sens.methods

## $aREAmir
## original       20       30       40       50 
##    0.875    0.750    0.750    0.604    0.417 
## 
## $EN.up
## original       20       30       40       50 
##    0.125    0.042    0.021    0.021    0.042 
## 
## $EN.down
## original       20       30       40       50 
##        1        1        1        1        1 
## 
## $EN.combined
## original       20       30       40       50 
##        1        1        1        1        1 
## 
## $wEN.up
## original       20       30       40       50 
##        0        0        0        0        0 
## 
## $wEN.down
## original       20       30       40       50 
##        1        1        1        1        1 
## 
## $wEN.combined
## original       20       30       40       50 
##        1        1        1        1        1 
## 
## $michael.up
## original       20       30       40       50 
##        0        0        0        0        0 
## 
## $michael.down
## original       20       30       40       50 
##    1.000    1.000    1.000    0.958    0.875 
## 
## $michael.combined
## original       20       30       40       50 
##    1.000    1.000    1.000    0.958    0.917 
## 
## $MW
## original       20       30       40       50 
##    1.000    1.000    1.000    1.000    0.896 
## 
## $KS
## original       20       30       40       50 
##    1.000    1.000    1.000    1.000    0.917 
## 
## $KS2
## original       20       30       40       50 
##    0.938    0.917    0.896    0.771    0.688 
## 
## $GSEA
## original       20       30       40       50 
##    0.125    0.125    0.125    0.125    0.125 
## 
## $GSEAmodified
## original       20       30       40       50 
##    0.438    0.438    0.438    0.438    0.479 
## 
## $modSites
## original       20       30       40       50 
##    0.938    0.938    0.917    0.896    0.875 
## 
## $modScore
## original       20       30       40       50 
##    1.000    1.000    0.958    0.979    0.854